]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
staging/rdma/hfi1: Prevent silent data corruption with user SDMA
authorMitko Haralanov <mitko.haralanov@intel.com>
Mon, 26 Oct 2015 14:28:37 +0000 (10:28 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 27 Oct 2015 08:19:22 +0000 (17:19 +0900)
User SDMA keeps track of progress into the submitted IO vectors by tracking an
offset into the vectors when packets are submitted. This offset is updated
after a successful submission of a txreq to the SDMA engine.

The same offset was used when determining whether an IO vector should be
'freed' (pages unpinned) in the SDMA callback functions.

This was causing a silent data corruption in big jobs (> 2 nodes, 120 ranks
each) on the receive side because the send side was mistakenly unpinning the
vector pages before the HW has processed all descriptors referencing the
vector.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/staging/rdma/hfi1/user_sdma.c

index 10ab8073d3f2453793da5c33d5daff1dca8500a9..36c838dcf0235790602f924ca490987899642514 100644 (file)
@@ -146,7 +146,8 @@ MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 12
 #define KDETH_OM_MAX_SIZE  (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
 
 /* Last packet in the request */
-#define USER_SDMA_TXREQ_FLAGS_LAST_PKT   (1 << 0)
+#define TXREQ_FLAGS_REQ_LAST_PKT   (1 << 0)
+#define TXREQ_FLAGS_IOVEC_LAST_PKT (1 << 0)
 
 #define SDMA_REQ_IN_USE     0
 #define SDMA_REQ_FOR_THREAD 1
@@ -249,13 +250,22 @@ struct user_sdma_request {
        unsigned long flags;
 };
 
+/*
+ * A single txreq could span up to 3 physical pages when the MTU
+ * is sufficiently large (> 4K). Each of the IOV pointers also
+ * needs it's own set of flags so the vector has been handled
+ * independently of each other.
+ */
 struct user_sdma_txreq {
        /* Packet header for the txreq */
        struct hfi1_pkt_header hdr;
        struct sdma_txreq txreq;
        struct user_sdma_request *req;
-       struct user_sdma_iovec *iovec1;
-       struct user_sdma_iovec *iovec2;
+       struct {
+               struct user_sdma_iovec *vec;
+               u8 flags;
+       } iovecs[3];
+       int idx;
        u16 flags;
        unsigned busycount;
        u64 seqnum;
@@ -294,21 +304,6 @@ static int defer_packet_queue(
        unsigned seq);
 static void activate_packet_queue(struct iowait *, int);
 
-static inline int iovec_may_free(struct user_sdma_iovec *iovec,
-                                      void (*free)(struct user_sdma_iovec *))
-{
-       if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
-               free(iovec);
-               return 1;
-       }
-       return 0;
-}
-
-static inline void iovec_set_complete(struct user_sdma_iovec *iovec)
-{
-       iovec->offset = iovec->iov.iov_len;
-}
-
 static int defer_packet_queue(
        struct sdma_engine *sde,
        struct iowait *wait,
@@ -825,11 +820,11 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                tx->flags = 0;
                tx->req = req;
                tx->busycount = 0;
-               tx->iovec1 = NULL;
-               tx->iovec2 = NULL;
+               tx->idx = -1;
+               memset(tx->iovecs, 0, sizeof(tx->iovecs));
 
                if (req->seqnum == req->info.npkts - 1)
-                       tx->flags |= USER_SDMA_TXREQ_FLAGS_LAST_PKT;
+                       tx->flags |= TXREQ_FLAGS_REQ_LAST_PKT;
 
                /*
                 * Calculate the payload size - this is min of the fragment
@@ -858,7 +853,7 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                                        goto free_tx;
                        }
 
-                       tx->iovec1 = iovec;
+                       tx->iovecs[++tx->idx].vec = iovec;
                        datalen = compute_data_length(req, tx);
                        if (!datalen) {
                                SDMA_DBG(req,
@@ -948,10 +943,17 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                                              iovec->pages[pageidx],
                                              offset, len);
                        if (ret) {
+                               int i;
+
                                dd_dev_err(pq->dd,
                                           "SDMA txreq add page failed %d\n",
                                           ret);
-                               iovec_set_complete(iovec);
+                               /* Mark all assigned vectors as complete so they
+                                * are unpinned in the callback. */
+                               for (i = tx->idx; i >= 0; i--) {
+                                       tx->iovecs[i].flags |=
+                                               TXREQ_FLAGS_IOVEC_LAST_PKT;
+                               }
                                goto free_txreq;
                        }
                        iov_offset += len;
@@ -959,8 +961,11 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                        data_sent += len;
                        if (unlikely(queued < datalen &&
                                     pageidx == iovec->npages &&
-                                    req->iov_idx < req->data_iovs - 1)) {
+                                    req->iov_idx < req->data_iovs - 1 &&
+                                    tx->idx < ARRAY_SIZE(tx->iovecs))) {
                                iovec->offset += iov_offset;
+                               tx->iovecs[tx->idx].flags |=
+                                       TXREQ_FLAGS_IOVEC_LAST_PKT;
                                iovec = &req->iovs[++req->iov_idx];
                                if (!iovec->pages) {
                                        ret = pin_vector_pages(req, iovec);
@@ -968,8 +973,7 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                                                goto free_txreq;
                                }
                                iov_offset = 0;
-                               tx->iovec2 = iovec;
-
+                               tx->iovecs[++tx->idx].vec = iovec;
                        }
                }
                /*
@@ -981,11 +985,15 @@ static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
                        req->tidoffset += datalen;
                req->sent += data_sent;
                if (req->data_len) {
-                       if (tx->iovec1 && !tx->iovec2)
-                               tx->iovec1->offset += iov_offset;
-                       else if (tx->iovec2)
-                               tx->iovec2->offset += iov_offset;
+                       tx->iovecs[tx->idx].vec->offset += iov_offset;
+                       /* If we've reached the end of the io vector, mark it
+                        * so the callback can unpin the pages and free it. */
+                       if (tx->iovecs[tx->idx].vec->offset ==
+                           tx->iovecs[tx->idx].vec->iov.iov_len)
+                               tx->iovecs[tx->idx].flags |=
+                                       TXREQ_FLAGS_IOVEC_LAST_PKT;
                }
+
                /*
                 * It is important to increment this here as it is used to
                 * generate the BTH.PSN and, therefore, can't be bulk-updated
@@ -1214,7 +1222,7 @@ static int set_txreq_header(struct user_sdma_request *req,
                                req->seqnum));
 
        /* Set ACK request on last packet */
-       if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
+       if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
                hdr->bth[2] |= cpu_to_be32(1UL<<31);
 
        /* Set the new offset */
@@ -1246,7 +1254,7 @@ static int set_txreq_header(struct user_sdma_request *req,
                KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
                          EXP_TID_GET(tidval, IDX));
                /* Clear KDETH.SH only on the last packet */
-               if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
+               if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
                        KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
                /*
                 * Set the KDETH.OFFSET and KDETH.OM based on size of
@@ -1290,7 +1298,7 @@ static int set_txreq_header_ahg(struct user_sdma_request *req,
        /* BTH.PSN and BTH.A */
        val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
                (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
-       if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT))
+       if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
                val32 |= 1UL << 31;
        AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
        AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
@@ -1331,7 +1339,7 @@ static int set_txreq_header_ahg(struct user_sdma_request *req,
                val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
                                        (EXP_TID_GET(tidval, IDX) & 0x3ff));
                /* Clear KDETH.SH on last packet */
-               if (unlikely(tx->flags & USER_SDMA_TXREQ_FLAGS_LAST_PKT)) {
+               if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT)) {
                        val |= cpu_to_le16(KDETH_GET(hdr->kdeth.ver_tid_offset,
                                                                INTR) >> 16);
                        val &= cpu_to_le16(~(1U << 13));
@@ -1358,10 +1366,16 @@ static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status,
        if (unlikely(!req || !pq))
                return;
 
-       if (tx->iovec1)
-               iovec_may_free(tx->iovec1, unpin_vector_pages);
-       if (tx->iovec2)
-               iovec_may_free(tx->iovec2, unpin_vector_pages);
+       /* If we have any io vectors associated with this txreq,
+        * check whether they need to be 'freed'. */
+       if (tx->idx != -1) {
+               int i;
+
+               for (i = tx->idx; i >= 0; i--) {
+                       if (tx->iovecs[i].flags & TXREQ_FLAGS_IOVEC_LAST_PKT)
+                               unpin_vector_pages(tx->iovecs[i].vec);
+               }
+       }
 
        tx_seqnum = tx->seqnum;
        kmem_cache_free(pq->txreq_cache, tx);